Umberto Lupo

DiffPaSS improves pairing of interacting protein sequences using differentiable optimization, outperforming existing methods on benchmark datasets.

This paper introduces DiffPALM, a method to pair interacting protein sequences using masked language modeling, improving structure prediction of protein complexes.

The paper demonstrates that MSA Transformer can generate protein sequences that perform as well as natural sequences in homology, coevolution, and structure-based measures, outperforming Potts models in some cases.

The paper investigates how phylogenetic correlations affect structural contact inference from protein sequences and shows global methods like Potts models are more resilient to these effects.

This paper shows that protein language models using MSA column attentions capture phylogenetic relationships and distinguish coevolutionary signals from historical contingency.